FexTracker technical details
From Aegisub Wiki
The goal of this page is to describe how FexTracker works and what the parameters mean.
This is also to help TheFluff document it in the help file.
Contents |
[edit] Features
A feature is a point whose movement is tracked.
Features are determined automatically, and the generation of them can only be controlled through some parameters. Features are created at the points of the image with the highest eigenvalue of the correlation matrix of the pixels in the vicinity.
[edit] Min Distance Squared
The Min Squared Distance is one of the parameters that control feature generation. It controls the minimum distance between two features at generation time. The larger this is, the further apart features will be, and thus they will (hopefully) cluster less.
[edit] Max CPU per feature
This controls how many iterations of motion estimation will at most be given to a feature at least level in the image pyramid. Higher numbers will potentially give better precision.
[edit] Min Displacement
If a feature has moved less than this during a motion estimation iteration at a given level in the image pyramid, the feature is considered to be "stable" and estimation will be stopped short, even if there were more possible iterations given by the Max CPU parameter.
[edit] Max Residue
This controls the maximum actual difference between the vicinity of a feature in two adjacent frames. Even if a feature has been motion estimated successfully, if the difference between the current and next frame at the feature's determined position in the next frame is too big, the feature will be dropped. The Max Rasidue is expressed in average absolute difference between the areas in the two frames.
The Max Residue can generally be exceeded in three cases:
- The feature has been covered by another object
- The feature has changed too much in appearance (changing color?)
- There is a high level of noise in the video (because of the guassian image pyramid, noise shouldn't matter much however)
While increasing the Max Residue might have features "last longer", it will also mean a larger risk of mis-tracking.
[edit] Window X and Window Y
Size of the window around a point in the image, used for determining how big chance a point has to be chosen as a feature, as well as the window searched for motion estimation. Effectively, the window for motion estimation might be a bit smaller though due to the algorithm used.
Having a large window might provide better motion estimation, but might also negatively influence feature detection, as well as slow down things generally. A too small window will even more so provide bad feature detection and obviously provide worse motion estimation. (Except perhaps in the case of a scene with only slight motion, where a too large window might mean more mis-estimations.)
[edit] Search Range
Used to determine how fast the pyramid declines in size, in conjunction with the window size. Should influence the range motion estimation considers.
